PSR-NET
Chandrayaan-2 OHRC · permanent shadow recovery
000 reading the frame
Chandrayaan-2 · Orbiter High Resolution Camera
Not once in two billion years. Scroll — we are going there.
Begin descent
01 · Approach
The Moon's axis is tilted 1.54°. With almost no obliquity there is no polar summer: the sun stays pinned to the horizon and rakes the landscape at grazing incidence forever.
02 · Permanent shadow
Any depression deep enough to hide behind its own rim stays dark in every season. Below 110 K, cold enough to trap water ice for geological time — and we have almost no pictures of the ground.
03 · The orbiter
Twenty-five centimetres per pixel from a hundred kilometres up — the sharpest eye ever flown to the Moon. It was built to find the boulder that would tip a lander over, before the lander found it.
04 · Target
The reticle drops onto a crater floor inside an Artemis candidate region. The craft slews, locks, and opens the shutter on ground that has never been photographed in reflected sunlight, because there has never been any.
05 · Descent
Through the rim shadow, onto the floor.
06 · Surface
Its lamp reaches a few metres. Everything past that is lit only by sunlight that bounced off the crater wall above — about a thousandth of the light falling on the rim.
07 · The picture
A 256-metre patch of floor at 25 cm per pixel, twelve bits deep, and almost all of it sitting on the noise floor. The rest of this page is about getting the terrain back out of it.
Chandrayaan-2 · Orbiter High Resolution Camera
It is a 256-metre patch of the floor of a crater at the lunar south pole, at 25 cm per pixel. The sun has not risen on it in two billion years. Every photon in it bounced off a sunlit wall first.
Scroll. The frame is going to open up.
Where this happens
That single number is the whole story. With almost no obliquity there is no polar summer to speak of: near the poles the sun stays pinned to the horizon and rakes across the landscape at grazing incidence forever. Any depression deep enough to hide behind its own rim never gets lit — not in winter, not in summer, not once in the last two billion years.
These are cold traps. Surface temperatures inside them sit below 110 K, cold enough that water ice delivered by comets and micrometeorites cannot sublime away on geological timescales. It is the single most valuable real estate in the inner solar system, and we have almost no pictures of the ground.
Why the frame is nearly black
A PSR is defined by never receiving direct sunlight. Whatever reaches the floor has scattered off an illuminated wall above, or arrived as Earthshine, or come from starlight and the zodiacal glow. Published estimates put the secondary irradiance between 10−3 and 10−5 of direct solar, depending on how much sunlit wall the spot can see.
OHRC is a push-broom instrument on a spacecraft moving at 1.6 km/s. Integration time per line is fixed by ground speed and the along-track sample spacing — stare longer and the image smears. Time delay integration buys back a factor of 64 by shifting charge down the array in step with the ground, and that is the entire margin available.
At these irradiances a pixel collects a handful of electrons. Read noise is 28 e− RMS, about 1.2 DN. The scene's whole dynamic range inside the shadow is roughly 12 DN out of a 4095-count well. The information is there — it is simply the same size as the noise, which is a different problem from being absent.
Bars are logarithmic. The linear picture is a 500:1 contrast ratio, which is why a single global stretch cannot show both.
The sensor
Twenty-five centimetres per pixel from a 100 km orbit. It was flown to certify landing sites — to find the boulder that would tip a lander over before the lander found it. Inside a PSR it is being asked to do that job with a thousandth of the light it was designed around, which is precisely why the recovery problem is worth solving rather than working around.
Every figure below is used directly by the sensor model that generated the frames on this page, so the simulation cannot quietly disagree with the specification it claims to follow.
| Ground sample distance | 0.25 m at 100 km |
|---|---|
| Swath | 3.0 km |
| Spectral band | 400–750 nm, panchromatic |
| Detector | TDI CCD, 64 stages |
| Line time | 290 µs |
| Quantisation | 12 bit (0–4095 DN) |
| Full well | 95 000 e− |
| Read noise | 28 e− RMS ≈ 1.2 DN |
| Dark signal | 180 e−/s → 3.3 e− per integration |
| PRNU | 1.2 % |
| Along-track smear | 2.4 px, TDI rate mismatch |
Stage 00
Twelve-bit digital numbers, displayed across their true range. The column striping is the sensor's own: each image column reads through its own amplifier chain, and the offsets differ. The mottling is photon shot noise plus 28 e− of read noise. Somewhere under it is a crater floor.
Stage 01
Bias and dark current are added to every pixel whether or not a photon arrived, so they come off first. The offset is estimated from the mode of the frame's own low tail rather than a calibration table, which keeps the method usable on data whose ancillary files you do not have.
Why first: every later estimator assumes zero-mean noise about a physical signal.
Stage 02
The stripes are constant down each column; real terrain is not. Taking the column-wise median of the residual against a horizontally smoothed copy isolates the electronic offset and leaves vertical geology intact. Watch the vertical banding leave without the ridges going with it.
Why here: striping is structured noise. A denoiser tuned for white noise treats it as signal and preserves it.
Stage 03
A guided filter fits a local linear model of the image against itself. Where the local variance is pure noise the model collapses to the local mean; where it is a crater rim, the edge passes through untouched. Linear time in pixel count, and no gradient reversal at the boulder edges this whole exercise exists to protect.
Why before deblurring: Richardson–Lucy amplifies whatever noise you hand it, without bound.
Stage 04
Inside a PSR the light comes from whatever patch of sunlit wall a spot can see, so irradiance falls by more than a decade across one frame. Multi-scale retinex divides out the illumination field at three scales at once and leaves reflectance — the property of the rock rather than of the geometry.
This is the step that makes a single global stretch work at all.
Stage 05
The blur is not anyone's mistake. It is the optical point spread function convolved with 2.4 pixels of along-track smear, because a TDI line rate never exactly matches ground speed. Richardson–Lucy is the maximum-likelihood inverse for Poisson data — which is what a photon-starved frame is — and it cannot return negative radiance, so rims sharpen instead of ringing into ghost pairs.
Iterations are capped at 24. Past that it stops recovering structure and starts inventing it.
Stage 06
CLAHE equalises tile by tile so a bright patch of wall-lit floor cannot crush the shadowed half of the frame. The clip limit caps the transfer slope, which is the honesty parameter of the entire pipeline: without it, a tile of flat noise gets stretched into convincing texture.
Clip limit 2.6. Raise it and the picture gets prettier and less true.
Stage 07
Everything above is the deterministic chain — operators anyone can audit, each one doing a thing you can name. This stage is PSR-Net: 714 000 parameters trained on synthetic frames whose sensor was re-randomised on every single sample, so it never saw the same noise, blur, gain or column pattern twice.
Running here, in your browser, on the same array the chain got. —
On 24 held-out scenes it scores 20.28 dB against the chain's 17.47. Held out means scenes and noise seeds it never trained on — but from the same generator, so it proves the network inverts this calibrated degradation better than a fixed operator chain, not that it works on real OHRC frames.
Stage 08
Laplacian-of-Gaussian blobs across a scale octave, scored on the block-and-shadow pair rather than on brightness — a noise spike has no shadow. Then the part that matters: every candidate is re-tested against independent noise realisations of the same scene. Real blocks survive resampling because they are in the signal. Hallucinations do not.
— detections retained, — rejected as unstable. Ring size is scale; opacity is confidence.
Stage 09 · Result
The right-hand half is the paired target: the same scene under the same light, rendered with a noiseless, smear-free sensor. Nothing in the pipeline has ever seen it. It exists so the numbers underneath the viewport can be measured rather than asserted.
Drag the divider, or focus the frame and use the arrow keys.
The model
There is no ground truth inside a PSR. Nobody has ever photographed the floor of Shackleton in reflected sunlight, so there is no clean image to regress toward. The published approach — and the one taken here — inverts the problem: model the scene and the sensor precisely enough that you can generate as many exactly-paired examples as you need, then learn the inverse of a degradation you defined.
A NAFNet-style U-Net: three scales, depthwise convolution, SimpleGate, channel attention, no batch norm. 714 401 parameters — deliberately small. A frame whose entire dynamic range is a dozen DN does not support a large model, and capacity beyond what the data supports is capacity available for inventing detail. No diffusion prior and no adversarial head, because both produce plausible texture, and plausible is exactly what a landing-site product must not contain.
Synthetic pairs generated by the same renderer that made this page's frames: crater populations on a d−2.6 size-frequency slope, boulders on d−3, illumination from a single-bounce wall-scattering model, and the full sensor chain — shot noise, dark, read noise, PRNU, column offsets, TDI smear, 12-bit quantisation.
Charbonnier on radiance, plus a gradient term on the log domain so faint edges count as much as bright ones, plus a Poisson likelihood term that ties the output back to the photon statistics of the input. No adversarial term: a discriminator rewards convincing texture, which is the failure mode being designed against.
The final layer predicts per-pixel variance as well as radiance (heteroscedastic regression), and Monte-Carlo dropout at inference gives an epistemic band on top. Every delivered product carries a confidence raster; a pixel the model is unsure about is marked, not smoothed over.
Honest scope
Both halves run here, on your machine, against the same 12-bit array: the deterministic operator chain, and the trained network. PSR-Net is 714,401 parameters — small enough to be a 2.9 MB static file, which is the only reason a web page can honestly claim to be running the model rather than describing it. It was trained on synthetic frames whose sensor was re-randomised every sample; the generator, the training script and the checkpoint are all in this repository.
Checking for weights…
Ablation
Both methods get the same input, both are scored after the same affine fit to the
target, and the held-out scenes are degraded with fixed seeds so the comparison is
reproducible. Run python psr/train/evaluate.py to regenerate it.
| Method | PSNR (dB) | SSIM | Δ vs raw |
|---|---|---|---|
| No benchmark file yet — train the model to produce one. | |||
Each row removes one stage and re-runs everything else against the paired target. These figures are computed in your browser when this section scrolls into view — if you change the demo parameters further down, they change too.
| Configuration | PSNR (dB) | SSIM | CNR | Δ PSNR |
|---|---|---|---|---|
| Scroll here to run the ablation… | ||||
Removing radiometric correction changes PSNR by 0.00 dB. It should: the affine fit used before scoring absorbs any constant offset, and bias is a constant offset. The stage still matters — every downstream estimator assumes noise is zero-mean about a physical signal, and the detector's shadow test compares absolute levels — but this particular metric is blind to it by construction. A metric that cannot see a stage is a fact about the metric.
Removing deconvolution improves PSNR slightly while SSIM falls. That gap is the whole argument of this page in one row. Richardson–Lucy trades pixel-wise error for structural fidelity: it sharpens rims, which SSIM rewards and squared error punishes, because a sharpened edge that is half a pixel off costs more MSE than a blurred edge that is centred. If you optimise a PSR pipeline on PSNR alone you will delete the deconvolution and ship a blurrier product that scores better.
Validation
Degrade a known scene through the calibrated sensor model, recover it, and measure PSNR, SSIM and LPIPS against the original. This is the only place full-reference metrics are legitimate, and it tests the inverse of exactly the degradation you modelled — no more.
On actual OHRC data there is nothing to compare to, so the metrics change shape: contrast-to-noise ratio, gradient energy, and BRISQUE/NIQE-class no-reference scores. They cannot tell you the picture is correct. They can tell you it stopped being noise.
ShadowCam on KPLO images PSRs directly at 200× the sensitivity of a conventional camera. Where its coverage overlaps an enhanced OHRC frame, the recovered structure has to match a genuinely independent observation. This is the strongest external check available, and the one a reviewer should ask for.
Validation, continued
Recovered slopes and shadow directions must be consistent with LOLA altimetry and with the illumination geometry at the time of acquisition. A boulder whose shadow points the wrong way relative to the sunlit wall is a detection error, and this test finds it without any reference image at all.
Re-run detection across independent noise realisations of the same frame and keep only what recurs. This converts "is it real?" from an aesthetic judgement into a measurable frequency, and it is what produces the confidence value attached to every detection above.
Plant synthetic boulders of known size and position into real frames, run the pipeline blind, and measure what fraction come back and at what size limit. Borrowed from observational astronomy, where it is the standard way to calibrate a detection completeness curve.
A generative model asked to enhance a dark image will produce a beautiful, confident, entirely fictional boulder field. If that product certifies a landing site, the error is not aesthetic. Every design decision on this page — no adversarial loss, clipped contrast limits, capped deconvolution iterations, ensemble stability testing, per-pixel uncertainty — exists to make the pipeline prefer saying "I don't know" over saying something plausible.
Provenance
A reviewer's first question is which parts are data. Answering it before being asked is the difference between a demonstration and a claim.
| Element | Status | Source and treatment |
|---|---|---|
| South-pole topography | Measured | LOLA gridded DEM at 16 px/deg, NASA/GSFC SVS CGI Moon Kit. Resampled to a polar stereographic frame at 189 m/px and smoothed at 3 px, still four times finer than the source grid. |
| PSR extent map | Derived | Horizon marching on that DEM from six azimuths at a 1.54° sun. Computed by this project; agrees with published PSR maps in extent and location, not traced from them. |
| OHRC instrument parameters | Published | GSD, swath, TDI stages, line time, quantisation and band from the mission instrument description. Noise figures are representative of a TDI CCD of this class. |
| The 256 m crater-floor frame | Synthetic | No altimetry resolves terrain at 0.25 m. Craters drawn from a d−2.6 size-frequency law, boulders from d−3, both scaled to a polar crater floor. Statistics, not terrain. |
| Illumination inside the PSR | Modelled | Single-bounce wall scattering at 3×10−3 of direct, gated by sky visibility. A first-order radiosity stand-in, good to a factor of a few — the accuracy published irradiance estimates themselves carry. |
| Sensor degradation | Modelled | Shot noise, dark, read noise, PRNU, column offset and gain, TDI smear, 12-bit quantisation, applied in sensor order — blur before noise, because the optics smear photons before the detector counts them. |
| Every metric on this page | Computed live | PSNR, SSIM, CNR, sharpness and the ablation table run in your browser against the paired target. Nothing is quoted from a results table. |
| Trained network weights | Trained here | PSR-Net was trained on this project's own generator — 384 cached scenes, sensor re-randomised every sample — and exported to ONNX with a numerical check against the PyTorch graph before shipping. The page loads the weights and runs them; if they are absent it says so in the model section rather than pretending. Generator, training script, checkpoint and benchmark are all in psr/train. |
| The learned-vs-deterministic table | Measured | Produced by evaluate.py on scenes held out of training, with fixed degradation seeds, both methods scored identically. The page reads that file; it does not restate it. |
Try to break it
The controls degrade the frame further, beyond what the sensor model already did, and the pipeline re-runs on the result. The metrics underneath are recomputed against the same untouched target every time, so you can find the point where recovery stops working — and it does have one.
Push blur past 2.5 px σ with a few DN of extra noise and watch SSIM collapse while the detection count climbs. Those are not recovered boulders — they are Richardson–Lucy amplifying noise into blocks, and the ensemble stability test is the only thing standing between that texture and a hazard map. Drop the iterations to 8 and the invented blocks disappear along with some real ones. That trade is the entire engineering problem.
Where it goes
Chandrayaan-3 landed at 69.4°S. Artemis III candidate regions sit within 6° of the pole, and several have PSR floors inside the landing ellipse. A boulder-and-slope hazard map inside those shadows is currently a gap in the mission-planning chain, filled today by inference from LOLA altimetry at 5 m/px — twenty times coarser than OHRC.
Surface frost signatures, sublimation pits and blocky ejecta exposing buried volatiles are all morphological. They need images, not just spectra, and the places worth looking are exactly the places that are dark.
The full chain is roughly 1.2 GFLOP per megapixel and processes an OHRC strip in seconds on one GPU. Nothing needs to fly: it consumes the archived radiometrically corrected product and emits an enhanced raster plus a confidence raster in the same projection.